{"auto_groups":[],"data":[{"model_name":"glm-5.1","description":"GLM-5.1 是面向智能体工程的下一代旗舰模型，采用混合专家（MoE）架构，拥有 754B 参数。该模型编程能力显著增强，在 SWE-Bench Pro 上取得了领先水平，在 NL2Repo（仓库生成）和 Terminal-Bench 2.0（真实终端任务）等基准上大幅领先前代。GLM-5.1 专为长周期智能体任务设计，能够以更好的判断力处理模糊问题，拆解复杂任务、执行实验、分析结果，并在数百轮迭代和数千次工具调用中持续优化。","vendor_id":1,"quota_type":0,"model_ratio":3.63,"model_price":0,"owner_by":"","completion_ratio":4,"enable_groups":["官方_70","官方_100"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"len \u003c 32000 ? tier(\"short_ctx\", p * 6 + c * 24 + cr * 1.3) : tier(\"long_ctx\", p * 8 + c * 28 + cr * 2)","pricing_version":"5a90f2b86c08bd983a9a2e6d66c255f4eaef9c4bc934386d2b6ae84ef0ff1f1f"},{"model_name":"doubao-seedance-2.0","vendor_id":29,"quota_type":0,"model_ratio":23,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["官方_80"],"supported_endpoint_types":["openai"]},{"model_name":"glm-5.3","vendor_id":1,"quota_type":0,"model_ratio":4,"model_price":0,"owner_by":"","completion_ratio":3.5,"cache_ratio":0.25,"enable_groups":["官方_100","官方_95"],"supported_endpoint_types":["openai"]},{"model_name":"mimo-v2.5-pro","description":"MiMo-V2.5-Pro 是小米发布的最新旗舰模型。与前代模型相比，它在通用智能体能力、复杂软件工程以及长程任务等方面都有显著提升，在 ClawEval、GDPVal 和 SWE-bench Pro 等基准测试中均位列前茅。它能够独立且完全自主地完成需要人类专家耗时数天甚至数周的专业任务，涉及上千次工具调用。其高达 100 万 token 的上下文长度，非常适合集成到各种智能体框架中使用。","vendor_id":30,"quota_type":0,"model_ratio":3.5,"model_price":0,"owner_by":"","completion_ratio":3,"cache_ratio":0.2,"enable_groups":["官方_100"],"supported_endpoint_types":["openai"]},{"model_name":"kimi-k2.6","description":"Kimi K2.6 是月之暗面（Moonshot AI）开源的原生多模态智能体模型，在 HLE（含工具）、SWE-Bench Pro、BrowseComp 等多项主流榜单上取得开源最优水平。 模型采用 MoE 架构，总参数量 1T，激活参数 32B，支持 256K tokens 上下文，并通过 MoonViT 视觉编码器支持图像与视频输入。 K2.6 专为智能体场景优化：支持 4,000+ 次工具调用与超 12 小时持续自主执行；多智能体协作可扩展至 300 个并行子智能体 × 4,000 步，单次运行生成 100+ 文件；同时支持思考（Thinking）与即时（Instant）双推理模式，以及 Function Call 与多轮思维链保留能力","icon":"Kimi","vendor_id":6,"quota_type":0,"model_ratio":3.25,"model_price":0,"owner_by":"","completion_ratio":4.153846153846,"cache_ratio":0.2,"create_cache_ratio":1.25,"enable_groups":["官方_100"],"supported_endpoint_types":["openai"]},{"model_name":"kimi-k3","description":"Kimi K3 是 Kimi 迄今能力最强的旗舰模型，拥有 2.8 万亿参数，基于 KDA 混合线性注意力机制（Kimi Delta Attention）和注意力残差（Attention Residuals）技术构建，并拥有 100 万 token 上下文窗口。它是全球首个开源的 3 万亿级别模型，面向长程编程、知识工作和推理等前沿智能场景而设计。","quota_type":0,"model_ratio":10,"model_price":0,"owner_by":"","completion_ratio":5,"cache_ratio":0.1,"enable_groups":["官方_100"],"supported_endpoint_types":["openai"]},{"model_name":"bge-base-zh-v1.5","description":"一款基于BERT架构的中文语义向量模型，以轻量级参数实现了接近大模型的性能。它专门针对中文优化，有效改善了相似度分布问题，在C-MTEB等基准测试中表现优异。模型采用MIT协议开源，是兼顾性能与效率的理想选择。","vendor_id":28,"quota_type":0,"model_ratio":0.25,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["官方_70"],"supported_endpoint_types":["embeddings"]},{"model_name":"MiniMax-M3","description":"MiniMax-M3是一个开源旗舰模型，集前沿编程能力、1M超长上下文和原生多模态三大能力于一身，是全球首个具备此完整能力组合的开源模型。它基于自研的MSA稀疏注意力架构，在保证性能的同时，将1M上下文下每Token计算量降至上一代的1/20，并在预填充和解码阶段实现了超9倍和15倍的加速。在编程与智能体评测中表现突出，例如在SWE-Bench Pro上得分59.0%，超越GPT-5.5和Gemini 3.1 Pro，并在BrowseComp（83.5分）等基准上取得领先。","vendor_id":26,"quota_type":0,"model_ratio":2.1,"model_price":0,"owner_by":"","completion_ratio":4,"cache_ratio":0.2,"enable_groups":["官方_100","官方_95"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"len \u003c= 512000 ? tier(\"standard\", p * 2.1 + c * 8.4 + cr * 0.42) : tier(\"long_context\", p * 4.2 + c * 16.8 + cr * 0.84)"},{"model_name":"qwen3.8-max","vendor_id":6,"quota_type":0,"model_ratio":6,"model_price":0,"owner_by":"","completion_ratio":3,"cache_ratio":0.125,"create_cache_ratio":1.25,"enable_groups":["官方_100"],"supported_endpoint_types":["openai"]},{"model_name":"doubao-seedance-2-0-260128","description":"✨ Seedance 2.0 视频 正式上线！最高支持 15s！请使用 /v1/videos 异步调用\n\n价格为官方价格的97折","vendor_id":29,"quota_type":0,"model_ratio":23,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["火山方舟_video"],"supported_endpoint_types":["openai-video"]},{"model_name":"qwen-image-2.0-pro","description":"Qwen-Image-2.0系列满血版模型，实现了图片生成和图片编辑的融合；具备更专业的文字渲染1k token指令支持能力、更细腻的真实质感，细腻刻画写实场景、更强的语义遵循能力。满血版具备2.0系列最强的文字渲染能力和真实质感。","vendor_id":6,"quota_type":1,"model_ratio":0,"model_price":0.5,"owner_by":"","completion_ratio":0,"enable_groups":["官方_80"],"supported_endpoint_types":["image-generation"]},{"model_name":"wan2.7-image-pro","quota_type":1,"model_ratio":0,"model_price":0.5,"owner_by":"","completion_ratio":0,"enable_groups":["官方_80"],"supported_endpoint_types":["openai"]},{"model_name":"deepseek-v4-flash","description":"DeepSeek V4 Flash is an AI model provided by venice.","vendor_id":2,"quota_type":0,"model_ratio":1.5,"model_price":0,"owner_by":"","completion_ratio":3,"cache_ratio":0.033333333333,"enable_groups":["官方_100","官方_70","官方_95"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"(tier(\"base\", p * 3.0 + c * 9.0 + cr * 0.10)) * (((hour(\"UTC\") \u003e= 4 \u0026\u0026 hour(\"UTC\") \u003c 6) || (hour(\"UTC\") \u003e= 10 || hour(\"UTC\") \u003c 1)) ? 0.5 : 1)"},{"model_name":"bge-m3","description":"BGE-M3 是一个多功能、多语言、多粒度的文本嵌入模型。它支持三种常见的检索功能：密集检索、多向量检索和稀疏检索。该模型可以处理超过100种语言，并且能够处理从短句到长达8192个词元的长文档等不同粒度的输入。BGE-M3在多语言和跨语言检索任务中表现出色，在MIRACL 和 MKQA 等基准测试中取得了领先结果。它还具有处理长文档检索的能力，在 MLDR 和NarritiveQA 等数据集上展现了优秀性能。","vendor_id":28,"quota_type":0,"model_ratio":0.25,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["官方_70"],"supported_endpoint_types":["embeddings"]},{"model_name":"deepseek-v4-pro","description":"DeepSeek-V4-Pro 是 DeepSeek-V4 系列中的旗舰 MoE 语言模型，拥有 1.6T 总参数、49B 激活参数，原生支持 100 万 tokens 的超长上下文。该模型采用创新的混合注意力架构，结合压缩稀疏注意力（CSA）与高度压缩注意力（HCA），在 1M 上下文下仅需 DeepSeek-V3.2 的 27% 单 token 推理 FLOPs 和 10% KV 缓存。模型还引入流形约束超连接（mHC）增强层间信号传播稳定性，并采用 Muon 优化器加速收敛。DeepSeek-V4-Pro 在超过 32T 高质量多样化 tokens 上预训练，后训练采用“领域专家独立培养 + 在线策略蒸馏统一整合”的两阶段范式。其最大推理强度模式 DeepSeek-V4-Pro-Max 在编程基准上取得顶尖表现，并在推理与 Agentic 任务上大幅缩小与领先闭源模型的差距，是目前最强的开源模型之一，支持 Non-think、Think High、Think Max 三种推理强度模式。","vendor_id":2,"quota_type":0,"model_ratio":6,"model_price":0,"owner_by":"","completion_ratio":2.25,"cache_ratio":0.025,"enable_groups":["官方_95"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"(tier(\"base\", p * 9.0 + c * 27.0 + cr * 0.30)) * (((hour(\"UTC\") \u003e= 4 \u0026\u0026 hour(\"UTC\") \u003c 6) || (hour(\"UTC\") \u003e= 10 || hour(\"UTC\") \u003c 1)) ? 0.5 : 1)"},{"model_name":"glm-5","description":"GLM-5 是智谱推出的新一代大语言模型，专注于复杂系统工程和长周期 Agent 任务。模型规模从 GLM-4.5 的 355B 参数（32B 激活）扩展至 744B 参数（40B 激活），预训练数据从 23T 增加到 28.5T tokens。GLM-5 集成了 DeepSeek Sparse Attention（DSA），在保持长上下文能力的同时大幅降低部署成本。在推理、编程和 Agent 任务上，GLM-5 在开源模型中达到同类最佳水平","vendor_id":1,"quota_type":0,"model_ratio":37.5,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["官方_75"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"len \u003c 32000 ? tier(\"short_ctx\", p * 4 + c * 18 + cr * 1.3) : tier(\"long_ctx\", p * 6 + c * 22 + cr * 2)"},{"model_name":"qwen3.7-max","description":"Qwen3.7-Max是阿里巴巴于2026年5月20日发布的智能体旗舰模型。它是一款万亿级参数的MoE架构模型，拥有100万Token的超长上下文窗口。其核心定位是长周期自主执行的智能体（Agent），能够连续运行长达35小时，处理超过1000次工具调用且性能不下降。在编程、推理等多项权威评测中，它位列国产模型第一，编程能力接近全球顶尖水平。最新版本还增加了视觉模态理解能力，具备多模态交互功能。","vendor_id":6,"quota_type":0,"model_ratio":6,"model_price":0,"owner_by":"","completion_ratio":3,"cache_ratio":0.2,"create_cache_ratio":1.25,"enable_groups":["官方_100"],"supported_endpoint_types":["openai"]},{"model_name":"kimi-k2.7-code","description":"Kimi K2.7 Code 是 Moonshot AI 推出的面向代码任务的 agentic 模型，基于 Kimi K2.6 构建，在真实世界长程代码任务中有显著提升，可增强复杂软件工程工作流中的端到端任务完成能力，并将思考 token 使用量相比 Kimi K2.6 降低约 30%。该模型采用 MoE 架构，总参数量为 1T，激活参数量为 32B，支持 256K 上下文长度，并支持图像输入、视频输入、Thinking mode、Preserve Thinking 与多步工具调用","icon":"Kimi","vendor_id":6,"quota_type":0,"model_ratio":3.25,"model_price":0,"owner_by":"","completion_ratio":4.153846153846,"cache_ratio":0.2,"create_cache_ratio":1.25,"enable_groups":["官方_100"],"supported_endpoint_types":["openai"]},{"model_name":"glm-5.2","description":"GLM-5.2 是 Z.ai 最新的旗舰模型，面向长程任务场景，相比 GLM-5.1 在长程任务能力上有显著提升。该 753B 模型支持稳定的 1M-token 上下文，具备更强的编程能力，并支持多种 thinking effort levels，以在性能和延迟之间进行灵活平衡。GLM-5.2 引入 IndexShare 架构优化，在 1M 上下文长度下将 per-token FLOPs 降低 2.9 倍，同时改进 MTP layer 以支持 speculative decoding，使 acceptance length 最高提升 20%","vendor_id":1,"quota_type":0,"model_ratio":4,"model_price":0,"owner_by":"","completion_ratio":3.5,"cache_ratio":0.25,"enable_groups":["官方_95","官方_100"],"supported_endpoint_types":["openai"]},{"model_name":"MiniMax-M2.5","description":"MiniMax-M2.5 是 MiniMax 推出的最新大语言模型，通过在数十万个复杂真实环境中进行大规模强化学习训练。该模型采用 MoE 架构，拥有 2290 亿参数，在编程、智能体工具调用与搜索、办公场景等任务中达到业界领先水平，在 SWE-Bench Verified 上取得 80.2% 的成绩，推理速度比前代 M2.1 提升 37%。","icon":"Minimax","vendor_id":26,"quota_type":0,"model_ratio":1.05,"model_price":0,"owner_by":"","completion_ratio":4,"enable_groups":["官方_70"],"supported_endpoint_types":["openai"]},{"model_name":"qwen3.6-plus","description":"Qwen3.6-Plus是阿里于2026年4月2日发布的旗舰大模型。它采用MoE架构，拥有100万上下文窗口，主打智能体编程与原生多模态能力。在多项评测中编程表现超越2-3倍参数的同类模型，接近全球最强编程模型Claude系列。","vendor_id":6,"quota_type":0,"model_ratio":37.5,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["官方_100"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"p \u003c= 256000 ? tier(\"base\", p * 2 + c * 12 + cr * 0.2 + cc * 2.5) : tier(\"tier_2\", p * 8 + c * 48 + cr * 0.8 + cc * 10)"}],"group_ratio":{"官方_100":1,"官方_40":0.4,"官方_70":0.7,"官方_75":0.75,"官方_80":0.8,"官方_95":0.95,"火山方舟_video":0.97},"pricing_version":"a42d372ccf0b5dd13ecf71203521f9d2","success":true,"supported_endpoint":{"anthropic":{"path":"/v1/messages","method":"POST"},"embeddings":{"path":"/v1/embeddings","method":"POST"},"image-generation":{"path":"/v1/images/generations","method":"POST"},"openai":{"path":"/v1/chat/completions","method":"POST"},"openai-response":{"path":"/v1/responses","method":"POST"},"openai-video":{"path":"/v1/video/generations","method":"POST"}},"usable_group":{"官方_100":"","官方_40":"官方价4折","官方_70":"","官方_75":"","官方_80":"","官方_95":"","火山方舟_video":"火山方舟_video"},"vendors":[{"id":5,"name":"Moonshot","icon":"Moonshot"},{"id":8,"name":"Qiniu"},{"id":18,"name":"GitHub Copilot","icon":"GithubCopilot"},{"id":19,"name":"xAI","icon":"XAI"},{"id":20,"name":"Meta","icon":"Ollama"},{"id":22,"name":"LLM Gateway"},{"id":26,"name":"Minimax 稀宇科技","icon":"Minimax"},{"id":9,"name":"Vivgrid"},{"id":12,"name":"Tencent Coding Plan (China)"},{"id":13,"name":"Xiaomi Token Plan (Singapore)"},{"id":14,"name":"Scaleway"},{"id":15,"name":"Zhipu AI Coding Plan"},{"id":21,"name":"Vultr"},{"id":23,"name":"aliyun-bailian"},{"id":30,"name":"xiaomi","icon":"XiaomiMiMo"},{"id":1,"name":"智谱","icon":"Zhipu.Color"},{"id":2,"name":"DeepSeek","icon":"DeepSeek.Color"},{"id":3,"name":"百度","icon":"Wenxin.Color"},{"id":6,"name":"阿里巴巴","icon":"Qwen.Color"},{"id":10,"name":"Venice AI","icon":"VeniceAI"},{"id":11,"name":"QiHang"},{"id":17,"name":"Anthropic","icon":"Claude.Color"},{"id":28,"name":"智源研究院","icon":"BAAI"},{"id":4,"name":"OpenAI","icon":"OpenAI"},{"id":16,"name":"OpenCode Go"},{"id":24,"name":"DigitalOcean"},{"id":29,"name":"字节跳动","icon":"Doubao.Color"}]}